RoPeSLR
3D RoPE-driven sparse-low-rank attention for efficient Diffusion Transformers.
RoPeSLR decouples attention into sparse and low-rank components to improve the efficiency of Diffusion Transformers while retaining 3D positional structure.
3D RoPE-driven sparse-low-rank attention for efficient Diffusion Transformers.
RoPeSLR decouples attention into sparse and low-rank components to improve the efficiency of Diffusion Transformers while retaining 3D positional structure.